Learning multiple models for exploiting predictive heterogeneity in recommender systems

Clinton Jones, Joydeep Ghosh, Aayush SharmaHetRec@RecSys 2011 · DOI

TL;DR

This paper introduces LaD-BAE, a Bayesian approach to recommender systems that tackles "predictive heterogeneity" by learning multiple localized models instead of a single global one. This allows for more accurate and interpretable recommendations by capturing variations in how user and item attributes influence preferences across different groups.

What Problem Does This Paper Address?

Recommender systems, particularly those using collaborative filtering, often rely on historical user-item interactions (like ratings) to make predictions. While effective, these methods can struggle with "cold-start" problems and often overlook valuable "side-information" such as user demographics or item characteristics. Content-based approaches leverage this side-information but typically use a single, global model. The core problem this paper addresses is that a single global model might not be optimal when different groups of users or items behave differently. This "predictive heterogeneity" means that the importance of various attributes can vary significantly across these subgroups, leading to suboptimal recommendations. The authors propose a method to exploit this heterogeneity by learning multiple, localized models.

Methods

The paper introduces the Latent Dirichlet Attribute Aware Bayesian Affinity Estimation (LaD-BAE) model. This Bayesian approach builds upon Latent Dirichlet Allocation (LDA) concepts. Key aspects include:

  • Soft Co-clustering: Instead of hard partitioning users and items into distinct groups, LaD-BAE uses a "mixed-membership" approach where each user and item can belong to multiple clusters with varying probabilities. This is achieved through a graphical model that represents the generative process of entity attributes and dyad ratings.
  • Localized Predictive Models: For each "co-cluster" (a combination of user clusters and item clusters), LaD-BAE learns a separate predictive model. This allows the model to assign different weights to various features (user/item attributes) depending on the specific co-cluster, effectively capturing predictive heterogeneity.
  • Variational Inference: Due to the complexity of the model, a variational approach with a fully factorized mean-field approximation is used to derive updates for model parameters and cluster assignments. This is an Expectation-Maximization (EM) style algorithm.
  • Attribute Integration: The model directly incorporates user and item attributes into the localized predictive models, allowing for a more nuanced understanding of preferences.

Datasets

The paper evaluates LaD-BAE on two datasets:

  • Yahoo! Movies Dataset: A dataset containing movie ratings, user attributes (age, gender), and movie attributes (average rating, genre, awards, etc.).
  • HetRec 2011 MovieLens Dataset: A dataset with movie ratings and derived user (bias) and item (bias, Rotten Tomatoes scores, etc.) attributes.

Key Insights and Ideas

  • Predictive Heterogeneity is Key: The paper strongly suggests that simply using side-information in a global model isn't enough. The variation in how that side-information influences preferences across different user/item groups is crucial for improving recommendations.
  • Soft Co-clustering Offers Flexibility: The mixed-membership approach allows for more nuanced groupings than hard clustering, better reflecting the complex relationships in real-world data.
  • Interpretability Through Local Models: Learning separate models for co-clusters makes the recommendations more interpretable. For instance, one can understand why a particular user group might like a movie based on the specific features that are important to that group.
  • Feature Importance Varies by Cluster: The study highlights that features important for one co-cluster might be irrelevant or even detrimental for another, underscoring the need for localized models.

Key Results

  • Accuracy: LaD-BAE achieved RMSE (Root Mean Squared Error) values comparable to or slightly better than well-known models like PMF (Probabilistic Matrix Factorization) and global linear models, especially on the MovieLens dataset. On Yahoo! Movies, it performed well, even recovering from less informative global features.
  • Interpretability: The authors emphasize that LaD-BAE provides significantly greater interpretability. By examining the learned coefficients (β values) for different co-clusters, one can understand which attributes are driving recommendations for specific user-item groups. For example, they observed that in one user cluster, actor and director preferences were more important than user bias, while in another, user bias was less influential.
  • Actionability: The localized models offer actionable insights. For instance, understanding that a specific user cluster might dislike popularly rated movies (as seen in Yahoo! Movies) allows for more targeted recommendations or marketing strategies.

Limitations and Future Work

  • Seeding Sensitivity: The LaD-BAE model's performance can be sensitive to initial parameter seeding, necessitating averaging over multiple runs.
  • Feature Selection: The paper notes that the performance can be significantly impacted by the choice of features. Removing less informative features improved LaD-BAE's performance on the MovieLens dataset.
  • Scalability: While the approach is designed to be more scalable than some alternatives, further investigation into its scalability for extremely large datasets is implied.
  • Future Work: The authors suggest exploring additional sources of information like tag data and social network information, and further investigating feature selection strategies to optimize model performance.

Summary updated

BibTeX
@inproceedings{jones2011learning,
  author = {Clinton Jones and Joydeep Ghosh and Aayush Sharma},
  title = {{Learning multiple models for exploiting predictive heterogeneity in recommender systems}},
  booktitle = {HetRec@RecSys},
  year = {2011},
  doi = {10.1145/2039320.2039323},
}
Figure